Cloudflare AI Search goes GA with image search and OCR

Cloudflare made AI Search generally available on October 1, 2026. AI Search is a managed service that indexes a company's documents and lets apps and AI agents search them. The general-availability release adds image search, text recognition for scanned PDFs and larger files, and it puts prices on the service. Billing starts on November 1.

For developers, the pitch is less plumbing. Building search over your own data usually means stitching together storage, an embedding model, a vector database and a ranking step. Cloudflare sells all of it as one product.

What AI Search is

AI Search combines four Cloudflare products into one index and retrieval pipeline, according to Cloudflare's announcement. Workers AI runs the models, Vectorize stores the vectors and R2 holds the files. The fourth is Browser Run, Cloudflare's hosted browser.

A vector, in this context, is a list of numbers that captures what a piece of content means. Searching by vectors finds results that are close in meaning, not just ones that share keywords. Cloudflare says it uses AI Search for its own blog and developer documentation search.

What is new at general availability

The biggest addition is searching images by what they show. AI Search now creates image embeddings, vectors made from the pixels themselves, using the Qwen3-VL-Embedding model. insideAI News describes this as search by the image's own pixels.

Feature	Change at GA
Images	Native image embeddings for visual search, via Qwen3-VL-Embedding
Scanned documents	OCR, or optical character recognition, reads text in scanned PDFs
File size	Text files and PDFs up to 10 MiB, up from 4 MiB
Embeddings	Matryoshka Representation Learning keeps vectors compact
Search modes	Hybrid search runs vector and keyword search in parallel and merges the results

Matryoshka Representation Learning trains embeddings so that a shorter version of each vector still works. That lets a service store and compare smaller vectors without losing much accuracy.

What it costs

Billing begins on November 1, 2026. insideAI News reports that parsing, chunking, embedding, keyword indexing and reranking come at no extra cost. Cloudflare's post lists these prices and monthly free amounts:

Item	Free each month	Price after that
Ingestion	5 million tokens	$0.75 per million tokens
Image ingestion	-	$0.50 per million tokens
Storage	10 GB	$2.00 per GB per month
Semantic queries	1,000	$0.75 per 1,000
Full-text queries	1,000	$0.10 per 1,000

Semantic queries search by meaning, using vectors. Full-text queries match keywords, and they cost much less.

Cloudflare has been building out its tools for AI agents. In August it showed Kitesurf, a browser built for AI agents that runs on its Workers platform.

What this means for developers

Price your queries before November 1. Semantic queries cost $0.75 per 1,000 after the free 1,000 a month, while keyword queries cost $0.10 per 1,000. An app that sends every search through the semantic path will cost more than seven times as much as one that uses keyword search where it is enough. Route simple lookups to full-text search.

Re-index your PDFs and images. If your index was built under the old 4 MiB limit, or skipped scanned documents, larger files and OCR may now surface content that was missing. Image embeddings also make product photos, screenshots and charts searchable by what they show.

Test hybrid search on your own queries. Merging vector and keyword results usually helps with names, part numbers and error codes that pure vector search misses. Compare results on a sample of real searches before you switch.

Watch storage as your index grows. The free tier covers 10 GB, then storage costs $2.00 per GB per month. Image-heavy indexes can grow quickly, so set a budget alert.

Weigh the lock-in. AI Search is convenient because it ties Workers AI, Vectorize, R2 and Browser Run together. That same tie makes it harder to move. Keep your source files in a format you could index elsewhere.